Orchestration

Ra conducts the orchestra. Fifteen AI models route through one pipeline. You brief once. Ra routes for the outcome you specified: watch-time, conversions, impressions, signups.

Orchestration is the thin harness. It is not where the intelligence lives. The intelligence lives in the skills (the five Intelligences, the templates, the playbooks) and in the deterministic tools (renderers, uploaders, schedulers). The harness routes. That separation is the whole architecture.

A fat harness tries to generate scripts and compose shots and optimize for virality and schedule posts inside one mega-prompt. It produces C+ results at everything. A thin harness picks the right skill for the task and hands it clean context. That is what Ra does, and that is why the same brief produces an A+ result across fifteen models instead of a mediocre result from one.

What Ra orchestrates

Ra sits above the model layer and the intelligence layer, picking the right specialist for each task.

Fifteen AI models in the stack

Primitive	Models Ra routes between
Motion / video generation	Veo, Kling, Runway, Pika, Luma, Seedance, Hailuo, MINIMAX
Keyframes / image generation	FLUX, Seedance, Midjourney, Ideogram
Dialogue / voice synthesis	ElevenLabs, Cartesia, OpenAI TTS
Score and SFX	Suno, Udio, generative SFX layers
Captions and dubbing	Whisper-class transcription plus multilingual TTS

Five Intelligences in the skill layer

• Media Intelligence — script generation, narrative structure, topic selection
• Cinematic Intelligence — shot composition, pacing, visual language
• Viral Intelligence — hook formulas, retention shaping, platform-native patterning
• Broadcast Intelligence — distribution, scheduling, cross-platform variants
• AdX Intelligence — conversion optimization, CTA placement, funnel alignment

Ra picks the right Intelligence for the creative decision, the right model for each shot, and the right deterministic tool for execution. One brief in. One coherent story out.

Routing decisions: latent vs deterministic

Every step Ra takes is either judgment (latent) or computation (deterministic). Confusing the two is the most common AI architecture mistake. Ra doesn’t make it.

Decision	Type	How it’s made
Which model for this shot’s mood and motion	Latent	Ra’s judgment, weighted by your Memory Graph
Hook variant for this audience	Latent	Viral Intelligence
Pacing tweaks between beats	Latent	Cinematic Intelligence
Whether to retake scene 3	Latent	Ra’s judgment plus your past retake signal
Aspect ratio per platform	Deterministic	Computed (9:16 for Shorts, 16:9 for YouTube, 1:1 or 4:5 for feed)
Duration limits per platform	Deterministic	Computed from platform specs
Credit budget per render	Deterministic	Computed from plan + project cap
Upload destination, filename, metadata	Deterministic	SDK calls

At CS scale, the deterministic layer handles roughly eighty percent of the work. The intelligence layer only fires for genuine creative decisions. That is why Ra is fast, cheap, and doesn’t hallucinate aspect ratios.

One brief, fifteen models

You write a Director Brief. Ra does the rest.

1. Ra reads the brief, your character sheets, your back catalog, and your outcome target.
2. Ra breaks the story into shots and assigns each shot to the model best suited for that shot’s subject, motion, mood, and your audience’s retention history.
3. Deterministic tools render, stitch, caption, score, and upload.
4. Outputs compose into one coherent story. Same color grade. Same voice. Same pacing. Same character across every model boundary.

You never subscribe to Runway and Pika and ElevenLabs and Kling separately. You never pick a model. You never guess which one is best this week. Ra knows, because Ra is routing thousands of shots per day across all of them and learning which model wins for which kind of shot.

The honest moat note

Ra’s routing is a mechanism, not a moat. Once competing orchestration layers ship, the routing itself is table stakes.

The moat is the Director Memory Graph and the Conversion Signal Flywheel. No model provider ever sees which hooks held past three seconds, which CTAs converted, which thumbnails earned impressions, which retakes you requested. CS sees all of it. That outcome data enriches the Graph. The Graph sharpens Ra’s next routing decision.

Graph is the flywheel. Ra is the turbine.

Ra’s value, long term, is not the routing. It is being the front door to a Memory that compounds with every story you ship.

Ra as IP, not feature

A feature (“AI agent that routes models”) is copyable. A named character with opinions, a social following, and a recognizable voice is brand IP. Ra is a director with taste, not a chatbot. Runway and Pika structurally cannot ship that.

Works with

Intelligence

The five Intelligences are the fat skills Ra routes to. Media, Cinematic, Viral, Broadcast, AdX. Each is a specialist. Read more

Memory

The Director Memory Graph is what Ra reads before every routing decision. Your thousandth scene ships cheaper, faster, and more on-brand than your first. Read more